Scholarlink

<p>Reliable tea quality evaluation remains challenging because existing visual or statistical models fail to capture the causal relationships among processing conditions, chemical composition, and sensory perception. To address this gap, this study proposes a causal deep-learning framework that integrates a Structural Causal Model (SCM) and a Counterfactual Generative Adversarial Network (CGAN) to model the production–composition–perception chain of tea quality. The SCM establishes causal dependencies among processing, physicochemical, and sensory variables, while the CGAN generates counterfactual representations to explain how processing changes influence quality outcomes. An attention-based interpretability module quantifies the contribution of each variable for transparent reasoning. Experiments on ISO 3720 and GB/T standard datasets demonstrate a 9.4% improvement in prediction accuracy and a 12.7% gain in interpretability compared with existing methods. Deployed on an edge device, the framework performs real-time inference, providing a practical, interpretable, and standard-aligned solution for intelligent tea-quality monitoring and industrial evaluation.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Causal deep learning for tea quality assessment via counterfactual generation and interpretability verification

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Lianlian Cao,
  • Ming Chen,
  • Zilan Wang

摘要

Reliable tea quality evaluation remains challenging because existing visual or statistical models fail to capture the causal relationships among processing conditions, chemical composition, and sensory perception. To address this gap, this study proposes a causal deep-learning framework that integrates a Structural Causal Model (SCM) and a Counterfactual Generative Adversarial Network (CGAN) to model the production–composition–perception chain of tea quality. The SCM establishes causal dependencies among processing, physicochemical, and sensory variables, while the CGAN generates counterfactual representations to explain how processing changes influence quality outcomes. An attention-based interpretability module quantifies the contribution of each variable for transparent reasoning. Experiments on ISO 3720 and GB/T standard datasets demonstrate a 9.4% improvement in prediction accuracy and a 12.7% gain in interpretability compared with existing methods. Deployed on an edge device, the framework performs real-time inference, providing a practical, interpretable, and standard-aligned solution for intelligent tea-quality monitoring and industrial evaluation.

Graphical abstract